Scholarlink

<p>Edge deployment of object detection faces a fundamental trade-off between computational efficiency and robustness to environmental variations. Existing lightweight detectors achieve real-time inference but suffer significant accuracy degradation under illumination changes, motion blur, and geometric distortions. This research proposes MAPBoost, a lightweight detection framework embedding augmentation resilience through three novel modules: SobelEdgePool for edge preservation, WeightedConcat for adaptive multi-scale fusion, and DepthMix for lightweight channel refinement. The synergistic integration of these modules creates cascading effects, redefining lightweight detection by achieving substantial improvements in both accuracy and efficiency. Evaluated on license plate detection—a challenging small-object scenario—MAPBoost achieves <b>38.1% improvement in mAP@75</b> over YOLOv8n while reducing computational cost by <b>56.8%</b> (3.5G vs 8.1G FLOPs). The framework maintains real-time performance (104 FPS on GPU, 40 FPS on edge devices) and demonstrates consistent robustness across environmental distortions. This novel architectural design establishes a new paradigm for edge-oriented detection, where robustness emerges from coordinated lightweight operations rather than computational overhead.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MAPBoost: augmentation-resilient real-time object detection for edge deployment

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Tu Nguyen Anh,
  • Vinh Dinh Nguyen

摘要

Edge deployment of object detection faces a fundamental trade-off between computational efficiency and robustness to environmental variations. Existing lightweight detectors achieve real-time inference but suffer significant accuracy degradation under illumination changes, motion blur, and geometric distortions. This research proposes MAPBoost, a lightweight detection framework embedding augmentation resilience through three novel modules: SobelEdgePool for edge preservation, WeightedConcat for adaptive multi-scale fusion, and DepthMix for lightweight channel refinement. The synergistic integration of these modules creates cascading effects, redefining lightweight detection by achieving substantial improvements in both accuracy and efficiency. Evaluated on license plate detection—a challenging small-object scenario—MAPBoost achieves 38.1% improvement in mAP@75 over YOLOv8n while reducing computational cost by 56.8% (3.5G vs 8.1G FLOPs). The framework maintains real-time performance (104 FPS on GPU, 40 FPS on edge devices) and demonstrates consistent robustness across environmental distortions. This novel architectural design establishes a new paradigm for edge-oriented detection, where robustness emerges from coordinated lightweight operations rather than computational overhead.